Documente Academic
Documente Profesional
Documente Cultură
Introduction
1
Department of Geology and Geophysics, University of Wisconsin-Madison, 1215 W. Dayton St.,
Madison, WI, 53706, USA. E-mail: hjzhang@geology.wisc.edu
374 H. Zhang and C. Thurber Pure appl. geophys.,
scale. As shown by synthetic tests (ZHANG and THURBER, 2003, 2005; ZHANG et al.,
2004), this method yields more accurate event locations and velocity structure near
the source region than standard tomography, which uses only absolute arrival times.
It has the unique ability to sharpen the velocity image near the source region because
of the combination of the higher accuracy of the differential time data and the
concentration of the corresponding model derivatives in the source region. The latter
results from the cancellation of model derivative terms where the ray paths overlap
away from the source region.
In many situations, CC data are not available due to the lack of waveform data.
There have been many applications using both DD location and tomography
algorithms demonstrating that using just the catalog pick differences leads to
significant improvement of event locations, as well as for the velocity structure for
DD tomography. As discussed in ZHANG and THURBER (2003) and ZHANG et al.
(2005), picking errors include two parts: random and systematic. Although random
errors may increase through the differencing process, the systematic errors will be
reduced or cancelled. The differencing process also causes the ray-path derivatives to
concentrate near the source region, rather than covering the entire model region. This
in turn reduces or removes the effect of the velocity model ambiguity outside the
source region on the source region model. As a result, the source region structure will
be better illuminated.
The DD tomography code tomoDD is built upon the double-difference location
code hypoDD written by WALDHAUSER (2001). In the original tomoDD algorithm, we
used an approximate pseudo-bending (ART-PB) ray-tracing algorithm (UM and
THURBER, 1987) to find the rays and calculate the travel times between events and
stations. The model is represented by velocity values specified on a regular set of 3-D
nodes and the velocity values are interpolated by using the linear B-spline
interpolation method. The hypocentral partial derivatives are calculated from the
direction of the ray and the local velocity at the source (LEE and STEWART, 1981).
The ray path is divided into a set of segments and the model partial derivatives
(calculated in terms of fractional slowness perturbation, so that the derivatives are
related to path length) are evaluated by apportioning the derivative to its eight
surrounding nodes according to their interpolation weights on the segment midpoint
(THURBER, 1983).
TomoDD assumes a flat earth model and is appropriate for local scale
problems (10’s to 100’s of kilometers). At the regional scale (100’s to 1000’s of
kilometers), however, sphericity of the earth should be taken into account. Major
velocity discontinuities such as Conrad, Moho, and subducting slab boundary
should also be considered. The ART-PB approach assumes a continuous velocity
model and cannot deal properly with velocity discontinuities. KOKETSU and
SEKINE (1998) have developed a spherical-earth 3-D ray-tracing method (including
discontinuities) using pseudo-bending, extended from a method developed in
Cartesian coordinates (ZHAO et al., 1992). ZHAO et al. (1992) and KOKETSU and
Vol. 163, 2006 Applications of Double-difference Seismic Tomography 375
Z k
Tki i
¼s þ u ds; ð1Þ
i
where si is the origin time of event i, u is the slowness field and ds is an element of
path length. For local earthquake tomography, the source coordinates ðx1 ; x2 ; x3 Þ;
origin times, ray paths, and the slowness field are the unknowns. The relationship
between the arrival time and the event location is nonlinear, thus a truncated Taylor
series expansion is generally used to linearize eq. (1). If we also discretize the velocity
model using a 3-D grid or mesh, we can write a linear equation relating the misfit
between the observed and predicted arrival times rki to the desired perturbations to
the hypocenter and velocity structure parameters
X
3
@T i Mik X
X N
rki ¼ k i
i Dxl þ Dsi þ wmn Dun Dsm ; ð2Þ
l¼1
@x l m¼1 n¼1
where Mik indicates the number of segments of the ray path from event i to station k
and wmn are interpolating weights of nth mesh node on the mid-point of the mth
segment with length Dsm . Subtracting a similar equation for event j observed at
station k from equation (2), we have
X
3
@T i Mik X
X N X
3
@T j
rki rkj ¼ k
i Dxil þ Dsi þ wmn Dun Dsm k
Dxjl Dsj
l¼1
@x l m¼1 n¼1 l¼1 @xjl
Mjk X
X N
wmn Dun Dsm ; ð3Þ
m¼1 n¼1
The observed differential arrival times ðTki Tkj Þobs can be calculated from both
waveform cross-correlation techniques for similar waveforms and absolute catalog
arrival times.
Note that the ray paths from two nearby events will substantially overlap,
meaning that the model derivative terms in eq. (3) will essentially cancel outside the
source region. For this reason, we include the absolute arrival times in the inversion
to resolve the velocity structure outside the source region. To combine absolute and
differential systems together, we apply a hierarchical weighting scheme during the
inversion to apply the appropriate relative weighting between them at different stages
of inversion. We start the inversion by applying higher weighting to the catalog data
(both differential and absolute catalog data) to establish the large-scale result (1 for
absolute data, 0.1 for differential catalog data, and 0.01 for cross-correlation data) in
Vol. 163, 2006 Applications of Double-difference Seismic Tomography 377
a manner similar to hypoDD (WALDHAUSER, 2001). Then for later iterations the
catalog differential data are weighted more to refine the event locations and the
velocity structure near the source regions (1 for differential catalog data, 0.1 for
absolute catalog data, and 0.01 for cross-correlation data). If waveform cross-
correlation data are available, they will be weighted more heavily than the catalog
differential data in the final iterations to further refine the event locations and the
velocity structure near the source region (1 for cross-correlation data, 0.01 for
differential catalog data, and 0.001 for absolute catalog data). We downweight the
differential catalog data by a factor of 100 in the final stage of the inversion, because
the cross-correlation data are one to two orders of magnitude more precise than the
manual picks (WALDHAUSER and ELLSWORTH, 2000). The relationship between
arrival time residuals and slowness model parameters is more linear than the event
locations (THURBER, 1992), as can be seen from eq. (1) where the integral in the
equation is linear in slowness but nonlinear in the hypocentral coordinates. This
indicates that the convergence of velocity structure is faster than event locations
(THURBER, 1992). Consequently, the simultaneous inversion of event locations and
velocity structure alternates with the inversion of only event locations in practice.
WOLFE (2002) carried out a strict and complete analysis of the use of difference
operators to relocate earthquakes. Next we conduct a similar analysis on using the
absolute data (i.e., standard location and tomography), the differential data (i.e., DD
location and tomography), and both of them to determine event locations and/or
velocity structure. Most of the notation used in this section follows WOLFE (2002).
Consider a set of p ¼ 1; . . . ; P earthquakes, with Np arrival times for each
earthquake. For each individual earthquake, eq. (2) can be represented in matrix
form as follows,
Ap DXp þ Cp DM ¼ DTp ; ð5Þ
For station corrections, we assume that each is a constant for all the arrivals
observed at a given station. As a result, let s0 stand for the station corrections for the
KT stations, which can be represented as
S ¼ Bs0 ; ð13Þ
where
2 3
B1
6 B2 7
B¼6 7
4 5; NT KT ; ð14Þ
BP
Let wk equal the number of earthquakes observed at station k, then the matrix QDD
P T wk ðwk 1Þ
has the dimension of Kk¼1 2 NT , corresponding to the possible combinations
of differences of earthquake arrival times at a station k, DTki DTkj (for i<j, to
eliminate repetition).
It can be seen that the difference operator QDD will annihilate the station
corrections as a result of QDD Bs0 ¼ 0. This is an advantage when comparing DD
tomography with standard tomography. The matrix form for the differential time
data of DD tomography is
QDD ADX þ QDD CDM ¼ QDD DT: ð17Þ
P KT
The difference operator QDD has the rank of k¼1 ðwk 1Þ, because the operation
removes one degree of freedom from the arrival times observed at one station
(WOLFE, 2002). This rank is usually considerably larger than the total number of
unknown parameters, including event hypocenter and origin time parameters and
model (slowness) parameters. If both matrices QDD A and QDD C have full column
rank, the DD tomography system will have the ability to determine absolute event
locations (MENKE and SCHAFF, 2004) and slowness (or velocity) structure uniquely.
Although the hypocenter partial derivatives for closely spaced events are very similar,
they will be quite different for events spaced far apart. Thus, in theory, if we include
all the possible combinations of pairs of arrival times observed at common stations,
the matrix QDD A should have full column rank. QDD C should also have full column
rank if we remove slowness model parameters not sampled by ray paths from the
inversion.
For the DD location algorithm in which the path anomaly biases between events
are not taken into account explicitly, only closely spaced events should be chosen in
order to reduce the path anomaly bias. Distance weighting is applied to reduce the
effect of event pairs that are far apart (WALDHAUSER and ELLSWORTH, 2000).
However, for closely spaced events, the partial derivatives will be very similar or
equal, thus the DD location algorithm is only capable of resolving relative locations
between events (WOLFE, 2002). For the DD tomography system, since the path
anomaly biases between event pairs are taken into account explicitly, we do not have
to limit the inversion only to the closely spaced events. Distance weighting is an
option in the DD tomography algorithm and is mainly designed to downweight
differential times resulting from far apart events based on the fact that the similarity
380 H. Zhang and C. Thurber Pure appl. geophys.,
of waveforms is less for far apart events, as are the common errors associated with
picks.
Equation (17) can be further transformed into a more compact form, as follows,
QDD EY ¼ QDD DT; ð18Þ
DX
where E ¼ ½ A C has the dimension of NT ðMT þ LÞ, and Y ¼ has the
DM
dimension of ðMT þ LÞ 1. The unknown parameters including hypocenters, origin
times, and slowness model parameters can be obtained, as follows,
The partial derivative operator used in DD location is QDD A. Note that a row of the
matrix ðQDD AÞ1 QDD can be used to indicate the dependence of the corresponding
model parameter (location or origin time in this case) on all the travel times.
The perturbations to event locations can also be obtained from absolute arrival
times by assuming the velocity structure is known and presumably correct,
where w is the relative weighting between absolute and relative arrival times, and I is
the identity matrix. The complete formula to solve the velocity structure simulta-
neously with event locations by using both absolute and differential data is
1
QDD E QDD
Y¼ DT: ð24Þ
wE wI
Vol. 163, 2006 Applications of Double-difference Seismic Tomography 381
For simplicity, we construct a 2-D synthetic model (Fig. 1) to show the properties of
these different systems of equations. There are 6 clusters of earthquakes (circles) and
each cluster has 4 closely spaced events with inter-event distances of less than 200 m.
In the X and Y directions, the distances between clusters are 10 and 15 km,
respectively. There are 36 stations around these events, with all the events observed at
each station. As a result, there are 894 travel times in total for all the event and
station pairs. The velocity is assumed to be constant at 5 km/s, thus the rays between
events and stations are all straight lines. The region is divided uniformly into 81 cells
(or blocks). The model derivatives relative to slowness are just the path lengths
crossing the cells.
Since this is relatively a small data set, we can use the singular value
decomposition method to calculate the singular values, model resolution and
50
1 2 3 4 5 6 7 8 9
45
10
40
19
35
28 Cluster 1 Cluster 4
30
Y (km)
37
25
46 Cluster 2 Cluster 5
20
55
15
64 Cluster 3 Cluster 6
10
73 81
5
5 10 15 20 25 30 35 40 45 50
X (km)
Figure 1
Map of stations (triangles) and earthquake clusters (circles) for the synthetic example. Each cluster has 4
closely spaced events, with interevent distances among them less than 200 m. The velocity model is
assumed to be a constant 5 km/s. The model region is divided into 81 cells uniformly, with numbers in the
cells indicating the model parameter indices.
382 H. Zhang and C. Thurber Pure appl. geophys.,
150
100
Abs . data only
50 Min. SV: 0.0761
0
0 20 40 60 80 100 120 140 160
400
Figure 2
The singular value distribution of the partial derivative matrix from using different data types to determine
event locations and velocity structure simultaneously.
uncertainties for the different systems (MENKE, 1989; ASTER et al., 2005). Figure 2
shows the singular values for the operators E, QDDE, (E, 0.1QDDE)T, and (0.01E,
QDDE)T when using the absolute data only, the differential data only, and both of
them (with relative weighting 1:0.1 and 0.01:1), respectively, to jointly determine
event locations and slowness structure. The system using the differential data to
jointly determine event locations and slowness structure has one zero singular value
(Fig. 2). The zero singular value is due to the fact that the partial derivatives relative
to the origin time for all events are equal and the differential operator has no ability
to recover the absolute origin times, only relative ones. Including the absolute data
with only a small weight is helpful to stabilize the system, by increasing the minimum
singular value to 0.0577 (Fig. 2). This indicates that including the absolute data into
the differential system gives it the ability to resolve the absolute origin times, as
would be expected. Also notice that the largest singular value from the system having
higher weighting on the differential data is about 2 times greater than that from the
system having higher weighting on the absolute data (Fig. 2). To resolve these
systems by using the damped least-squares algorithm, different damping values
should be chosen to make their condition numbers similar. The damping value that
makes the condition number of the system close to 60 to 80 is consistent with the
Vol. 163, 2006 Applications of Double-difference Seismic Tomography 383
Slowness
0.9 Origin
time
0.8
0.6 Location
diff
0.5
0.1*abs+diff
0.4 abs+0.1*diff
abs
0 50 100 150
Model parameters
Figure 3
Model resolution for location and slowness model parameters when using different data types to determine
event locations and velocity structure simultaneously by a damped least squares solution. The damping
values are 1.5 for the systems of abs and abs+0.1diff, and 4.4 for the systems of diff and diff+0.01abs. For
the model parameters, the slowness model cells are arranged at first in the sequence as shown in Figure 1,
followed by the location and origin time parameters.
one chosen based on the trade-off curve between data variance and model norm. For
this reason, we choose damping values of 1.5 and 4.4 for the systems having higher
weighting on the absolute data and higher weighting on the differential data, which
would make all the systems have condition numbers 70.
Figure 3 shows model resolution for origin time, location and slowness model
parameters. The systems with higher weighting on the differential data seem to better
resolve event locations by having higher resolution values (0.7 versus 0.3). For
slowness model parameters around the edge of the model, the systems having higher
weighting on the absolute data have higher model resolution. We also note that the
model resolution is generally higher for the slowness parameters than location
parameters. By carefully checking the singular values of the Fréchet derivative
matrices for slowness and locations parameters, we found that the largest singular
value for the slowness derivative matrix is considerably larger than that for the
location derivative matrix. For example, the largest singular value for the slowness
matrix in the case of using just differential times is 325.3, compared to 29.5 for the
location matrix. Since we use the same damping value of 4.4 for slowness and
location parameters, the latter ones are overdamped. This is the reason why the
model resolution for slowness parameters is greater than location parameters. This
overdamping for location parameters is justified since the relationship between
arrival time residuals and hypocenter perturbations is nonlinear. This is helpful for
384 H. Zhang and C. Thurber Pure appl. geophys.,
preventing large data residuals (or noise) from projecting into location parameters.
We also note that events are generally relocated one more time based on the final
velocity model.
We assume that the errors associated with both absolute and differential data are
an uncorrelated Gaussian-distributed random process. The standard deviations for
the absolute and differential data are 80 and 20 ms, respectively. This simulates the
scenario that the differential data are more accurate than the absolute data. These
values would be representative of absolute catalog picks and differential catalog
picks; using the even more precise CC differential times would obviously make the
error contrast greater. Figure 4 shows the model uncertainties (s/km for slowness
parameters and km for locations) for different systems. The model uncertainties
estimated from the system having higher weighting on the absolute data are
substantially larger, especially for event locations and origin times. For all the
systems, the model uncertainties associated with origin times are much smaller than
those for hypocenters. This is caused by the different dimensionality between the
origin time and hypocenter location (KLEIN, 1978). For example, a change of several
km in hypocenter location is equivalent to a one second change in origin time. We
also notice that the slowness model cells closer to two model region edges (X ¼ 5 and
50 km) tend to have larger model uncertainties. This phenomenon results from the
relatively fewer rays crossing through those cells. The characteristics of model
uncertainties having higher weighting on the differential data (diff and
0.04
diff
0.1*abs+diff
0.03 abs+0.1*diff Location
Model uncertainties
abs
Origin
0.02 time
Slowness
0.01
0
0 50 100 150
Model parameters
Figure 4
Model uncertainties for location and slowness model parameters when using different data types to
determine event locations and velocity structure simultaneously by a damped least squares solution. The
damping values for different systems are shown in Figure 3. The model parameters are arranged in the
same way as Figure 3.
Vol. 163, 2006 Applications of Double-difference Seismic Tomography 385
diff+0.01*abs) are quite similar, with the system (diff+0.01*abs) using the
combination of data types having slightly larger uncertainties for some slowness
model cells.
These results demonstrate the advantages of using both absolute and differential
data. The model resolution for location model parameters is higher in the systems
having higher weighting on the differential data. The system using both the absolute
and differential data with the relative weighting 0.01 to 1 between them has higher
model resolution and lower model uncertainties for location parameters than the
system using the absolute data only. For some model-edge slowness model cells, the
systems having higher weighting (abs and abs+0.1*diff) on the absolute data have
higher model resolution and lower model uncertainties.
straight lines, and the travel-times are calculated based on these straight lines. When
the grid interval is too large, the errors associated with this simplification will be
considerable. PODVIN and LECOMTE (1991) used a recursive scheme to reduce this
wavefront distortion to some extent. In comparison, the finite-difference scheme
(a) (c)
Air nodes
1 1
0.5 0.8
0
Z
0.6
Z
-0.5
0.4
-1
-1
-0.5 1 0.2
0 0.5
0.5 0
Y -0.5 X 0
1 -1 -0.5 0 0.5
X
(b)
0.8
Z
0.6
-0.5
0.5
0
Y
0
X
0.5
-0.5
Figure 5
Grid setup for regional scale double-difference tomography. (a) The whole earth is inserted into a cubic
box. (b) A rectangular box covers the region of interest. (c) The representation of (b) in the X-Z plane. The
surface of the Earth is shown as a thick line. A regular grid is used for the finite-difference travel-time
algorithm.
Vol. 163, 2006 Applications of Double-difference Seismic Tomography 387
proposed by HOLE and ZELT (1995) does not have such a scheme to deal with the
wavefront distortion near the source region.
Recently, ROECKER (pers. comm., 2003) proposed to calculate the travel-time field
in the source region with a much finer grid, and then combine this source region travel-
time field with that outside the source region. In this way, the effect of wavefront
curvature in the source region will be greatly reduced and the accuracy in the travel-time
field will also be improved without sacrificing too much computational time and
memory space. This strategy will be incorporated in future versions of tomoFDD.
To adapt DD tomography to the regional scale, the curvature of the Earth must be
taken into account. Following FLANAGAN et al. (2000), we solve this problem by
parameterizing a spherical surface inside a Cartesian volume of grid nodes. Figure 5a
shows the case of putting the Earth into a cube; this model can be used for global seismic
tomography. For regional seismic tomography, we are only interested in a part of the
Earth, so we construct a rectangular box covering the part of interest (Figs. 5b and c).
The coordinate center is placed at the surface of the Earth, the positive X and Y axes
point in the directions of North and West, respectively, and the positive Z axis points
downward. The grid nodes above the Earth’s surface (air nodes) are given the velocity
for P waves traveling in air (Fig. 5c). As a result, all the rays travel inside the Earth.
There are two different grids in the algorithm: one is the inversion grid that can be
nonuniform but regular and the other is the computational grid that must be
uniform. The uniform grid is used to calculate the travel-time field using the finite-
difference algorithms and the velocity values at its nodes are interpolated from the
non-uniform inversion grid through linear B-spline interpolation. If an inversion grid
node is 2 or more km above the Earth’s surface (allowing for topography), then it is
treated as an air node and its value is fixed during the inversion. First we treat each
station as a source and calculate travel times to all velocity nodes in the volume. The
travel time from a station to each earthquake is interpolated from its 8 neighboring
nodes through linear B-spline interpolation. The ray path from the earthquake to the
station is found iteratively with increments opposite to the travel-time gradient. After
these set-ups, the ray paths and the related partial derivatives are calculated by the
finite-difference method. The code has the option to choose the finite-difference
method either by PODVIN and LECOMTE (1991) or by Hole and ZELT (1995). In our
experience, the former method treats the source region problem better than the latter
method, but it calculates the travel-time field more slowly. Our future work is to
incorporate Roecker’s multi-grid method (pers. Comm., 2003) into the tomoFDD
algorithm to better deal with the source region problem.
The algorithms tomoDD and tomoFDD have already proven to be valuable tools
for tomography applications (see next section), but both of them are based on a
388 H. Zhang and C. Thurber Pure appl. geophys.,
of a tetrahedron when the sum of DWS values on its 4 nodes satisfies a predefined
threshold value (ZHANG and THURBER, 2005). The threshold checking is repeated to
determine the inversion mesh to be used for the current iteration of simultaneous
inversion (ZHANG and THURBER, 2005). Once the inversion mesh is determined to be
data adaptive, a new tetrahedral or Voronoi diagram is constructed and the partial
derivatives of the travel times with respect to the new set of inversion mesh nodes are
calculated using the saved rays for the construction of the seismic tomography
equations. After each simultaneous inversion, the velocity values on the irregular
inversion mesh nodes and the regular computational grid nodes are updated. For
subsequent simultaneous inversions, the inversion mesh is again updated following the
same procedure to better match with the ray distribution, which will change as the
velocity model changes and hypocenters move.
The adaptive mesh method considers the different ray distribution between the P
and S waves. Due to the relative simplification of picking P arrivals compared to S,
normally there will be more data and hence more inversion mesh nodes for P waves
than S waves. At each simultaneous inversion, the same procedure as discussed above
is used to construct different inversion meshes for P and S waves.
The above procedure can be applied directly for absolute data with the inversion
mesh built according to the absolute ray paths. In the case of using a combination of
absolute and differential data, however, the model derivatives from the differential
data are included in the mesh refinement process, so the inversion mesh will be
distributed more finely near the source region where these derivatives are largest. In
this case, we keep those irregular inversion mesh nodes outside the source region and
fix their velocity values if their DWS values are smaller than a threshold value
(ZHANG and THURBER, 2005). In this way, we preserve the velocity structure outside
the source region that is determined by the absolute data information while refining
the velocity structure near the source region.
Once the inversion mesh is set up, we can calculate the partial derivatives of travel
times with respect to the event locations and slowness model parameters. In a similar
way to the case of the regular inversion grid, we can then construct the linear system
of equations to solve for the event locations and slowness perturbations at the
irregular mesh nodes.
immediately deployed in and around the source region of the 2004 mid-Niigata
prefecture earthquake. Their results also showed the aftershocks were distributed
along a clear boundary between the low and high velocity structures corresponding
Figure 6
(a) Event and inversion grid distribution map of Northern California. Events are indicated by dots and
inversion grid nodes are indicated by crosses. Thin lines indicate fault traces. The coordinate center is
located at latitude 31.5 N and longitude 122 W. (b) NE-SW cross section of P-wave velocity model from
Pt. Ano Nueva in the southwest to near the town of Tracy in the northeast.
392 H. Zhang and C. Thurber Pure appl. geophys.,
to the hanging wall and footwall, respectively. TAKEDA et al. (2004) studied the
Atotsugawa fault in central Japan using tomoDD. They found a low-velocity zone up
to 10 km depth in the cross section along the fault and the earthquakes distributed
along the upper boundary of this low-velocity zone.
We also applied the adaptive-mesh version of DD tomography (tomoADD) to the
Parkfield, California data set (ZHANG and THURBER, 2005) using both linear and
natural neighbor interpolation methods. Both methods resulted in more clustered
earthquakes than standard tomography, and the events are concentrated where the
velocity contrast is sharpest. The velocity models using both interpolation methods
are characterized by a clear velocity contrast across the fault and they are
comparable in most areas, with the velocity model using natural-neighbor interpo-
lation being slightly smoother. We also found a velocity reversal at 4 km depth to
the southwest of the fault when using only the absolute arrival times, but this
anomaly disappears when including the differential data in the inversion. Synthetic
tests indicate that this velocity reversal could be caused by the noisier absolute data.
Thus including the differential data in the inversion may be helpful for removing
some velocity features that could be artifacts resulting from the noisier absolute data.
Our current work involves the development of a regional crustal model for
Northern California (Fig. 6a). The model will be used for improving earthquake
locations and for generating synthetic waveforms and estimating ground motions for
the 1906 San Francisco earthquake. Approximately 6000 earthquakes with arrival
data extracted from the Northern California Earthquake Data Center (NCEDC) and
about 5,500 shots or airgun blasts are used in this study. 466,368 absolute catalog
times and 1,629,973 differential times were used in the inversion. The inversion grid
was rotated so that the +X and +Y axes are approximately SAF-normal and SAF-
parallel, respectively. The grid intervals were 5 km in the X direction and 10 km in
the Y direction. In the Z direction, grid nodes were positioned at )1, 0, 1, 3, 6, 9, 12,
15, 20, 25 and 30 km. For this study, we used the tomoFDD algorithm considering
the relatively large size of the study region. An example cross section through a
preliminary model with 5 · 10 km horizontal gridding is shown in Figure 6b. The
section runs from Pt. Ano Nueva in the southwest to near the town of Tracy in the
northeast. At X = )25 km, we see the deep seismicity that is the northwestern edge
of the Loma Prieta rupture zone. Just to the northeast, at X = )20 km, is the diffuse,
shallower seismicity representing the San Andreas fault, marked by a distinct low-
velocity zone. Adjacent to that, at X = )15 km, the deep, low-velocity, Cupertino
Basin is evident, with an active fault situated beneath it. On the northeast side of the
section, the northeast-dipping Calaveras Fault seismicity is evident at X = +10 km,
associated with the edge of a dipping high-velocity block. There is earthquake
activity on both sides of the Calaveras here — the shallower activity to the southwest
appears to mark the beginning of the Hayward fault, whereas the deeper activity to
the northeast (X = +15 to +25 km) is on an obliquely striking fault connecting the
Calaveras and Greenville faults.
Vol. 163, 2006 Applications of Double-difference Seismic Tomography 393
-40
Y=
35
km
Y=
Y
0
Kapiti
km
Y=
Island
-3
CO
5
km
OK
-41
Latitude ( O S)
UGH
RO
G IT
AN
ST
R
KU
RA
GH HI
OU
IT
B OR
RL
X
-42 MA
km
0 50
could not resolve the detailed internal structure of the down-going slab using
standard tomography; their models showed the slab as a relatively homogeneous
high velocity anomaly. Our model, however, showed for the first time that the
interior of the slab actually has a highly heterogeneous velocity structure. The upper
plane of earthquakes lies in a region with relatively low Vp and Vs values and
average to high Vp/Vs ratios (1.72–1.85) that may correspond to the transformations
of metabasalt and metagabbro to blueschist. The lower plane of earthquakes exists in
a region of relatively low Vp and high Vs values and low Vp/Vs ratios that could be
due to dehydration reactions of serpentine and other hydrous minerals. The region
between the two planes of earthquakes has relatively high Vp/Vs ratios that indicate
partial hydration of the normal dry mantle. Synthetic tests with and without key slab
features found in this study using the same data distribution as the real data showed
the model was well resolved. SHELLY et al. (2004) found similar results when they
applied tomoDD and waveform cross correlation to the subduction zone near
Ibaraki, Japan that is located to the south of ZHANG et al. (2004)’s study region.
We further applied tomoFDD to study the transition zone from oblique
subduction to active continent-continent collision in the Wellington region, New
Zealand, where a double seismic zone is also present (Fig. 7). We used 5471
earthquakes spanning the transition region between the North and South Islands
that were recorded on 56 stations during the period 1990–2001. For the chosen events
and stations, there are 58,082 catalog P data, 299,948 P differential catalog data, and
427,771 WCC P differential times that were verified using a bispectrum analysis
method (DU et al., 2004). We adopt a coordinate center of latitude )41.3, longitude
174.8, and depth 0 km (with respect to sea level). The coordinate system is rotated in
the horizontal plane by 45 clockwise to be consistent with the strike of the
subducting slab. The inversion grid intervals are 10 km in the X direction and 15 to
20 km in the Y direction (Fig. 7). The node spacing in the vertical direction is
4.5 km. In addition to a low-velocity anomaly present in the upper part of the slab
corresponding to the subducting oceanic crust and sediments, our model also shows
varying features deeper inside the slab (Fig. 8). For the cross sections beneath the
North Island, the P-wave velocity is relatively high between the two planes, and it is
relatively low in the lower part of the slab, partly in association with the lower plane
of seismicity (Fig. 8). Similar to the subducting slab beneath northern Honshu, Japan
(ZHANG et al., 2004), the low Vp region associated with the lower plane of seismicity
might be caused by dehydration reactions of serpentine. However, the features are
quite different between the cross sections beneath the North versus the South Island
(Fig. 8). In the South Island cross sections, the velocity is relatively low and there is
almost no seismicity between the two planes of seismicity (Fig. 8). The lower plane of
seismicity is associated with a relatively high velocity region (Fig. 8). One possible
cause of such features is that the water released from the dehydration reactions of
serpentine escaped from the lower plane and entered into the region between the two
planes. As a result, the minerals in the lower plane may be mainly composed of
Vol. 163, 2006 Applications of Double-difference Seismic Tomography 395
Figure 8
NW-SE cross sections of the New Zealand Vp structure at Y=)35, 0, and 35 km. Structure above the
dashed line is meaningful, where the DWS for each node exceeds 15.
forsterite and enstatite, which have relatively high Vp (JI and WANG, 1999). The
water released into the region between the two planes decreases Vp. The different
features may be related to the transition from subduction to a strike-slip plate
boundary from northeast to southwest, and stagnation of the subducting slab in the
southwest.
We also found velocity variations in and around the subducting slab for the
Kodiak, Kenai, and McKinley blocks for the Alaska subduction zone using
396 H. Zhang and C. Thurber Pure appl. geophys.,
tomoFDD (ZHANG et al., 2004). For the Kodiak block, a low velocity zone is present
within the slab right below the high velocity zone where most of the Wadati-Benioff
Zone (WBZ) earthquakes are located. In comparison, for the Kenai block, the
velocity is higher in the zone below the WBZ earthquakes. The McKinley block
shows a more complex slab velocity pattern, with high and low velocity zones at
different depths within the slab. The high-resolution velocity structure may provide
additional constraints for models of the segmentation of the Alaska subduction zone
(RATCHKOVSKI and HANSEN, 2002), in addition to earthquake locations, volcanic arc
geometry and composition, etc., and may provide a possible explanation for the
segmentation.
Another application of tomoFDD is to image crust, mantle, and slab structure
around the Shikoku district, Japan, and to study its relationship to the occurrence of
nonvolcanic deep tremors (NAKAJIMA et al., 2004). High-velocity zones are associ-
ated with the deep seismic zone beneath the western part of Shikoku and the Kii
Channel that may indicate the presence of the subducting Philippine Sea plate. Low
velocity zones lie immediately above them in the western part of Shikoku and the
central part of the Kii Peninsula, and the Vp/Vs ratios are greater than 1.8. The
nonvolcanic deep, low-frequency tremors occurring along the strike of the subduct-
ing Philippine Sea detected by OBARA (2002) occur right in and around the low-
velocity and high Vp/Vs zones, lending support to the idea that fluids are involved in
the tremor. This study, along with those discussed above, shows strong evidence for
dehydration reactions in the subducting slab.
Thanks to the ability of DD tomography to resolve the finer structure near the
source region and the seismicity distribution inside the subducting slab, it is now
possible to characterize the fine details of the velocity structure and accurate
earthquake locations inside the slab, as shown in studies discussed above. This will in
turn be extremely helpful for understanding the constitution of the slab, the cause of
the intermediate depth earthquakes inside the slab, the fluid distribution and
recycling, and tremor occurrence (HACKER et al., 2001; OBARA, 2002).
Figure 9
(a) Event and station map for Mt. Etna. The summit crater is located at latitude 37.75N and longitude
15.0E. (b) Horizontal section of P-wave velocity model at depth of 1 km with earthquakes within 0.5 km
plotted as dots. (c) W-E cross section of P-wave velocity model at Y=)2 km determined using both
absolute and differential data. (d) W-E cross section of P-wave velocity model at Y=)2 km using only the
absolute data. For (c) and (d), earthquakes within 1 km of the cross section are plotted as dots.
398 H. Zhang and C. Thurber Pure appl. geophys.,
times. The inversion grid intervals in the horizontal and vertical directions are 2 km
and 1 km, respectively. In the horizontal section at depth of 1 km (Fig. 9b), a high
P-wave velocity zone of 4.8 – 5.1 km/s exists beneath the southeastern part of the
summit area, similar to the results of PATANÉ et al. (2002). This high-Vp body
extends to a depth up to 15 km as shown in the W-E cross-section of Y = )2 km in
the velocity model (Fig. 9c), and is interpreted as a mainly solidified intrusive body
(PATANÉ et al., 2003).
For comparison, Figure 9d shows the same W-E cross section of Y = )2 km
from the velocity model calculated using only absolute data. The velocity model is
more sharply defined in the inversion using both absolute and differential times. In
particular, the high velocity body shows a greater velocity contrast and more detailed
features (Fig. 9c) than in the absolute-data-only model (Fig. 9d). Compared to the
results using only absolute data (Fig. 9d), the event locations determined using both
absolute and differential data are more clustered. Two clusters of earthquakes are
present right above the high velocity body from depths of 1 km above sea level to
3 km below sea level. The western cluster is more concentrated than the
southeastern cluster and is situated on the edge of a knob of high velocity (Fig. 9c),
which is not evident in the velocity model using only absolute data (Fig. 9d; PATANÉ
et al. 2002). This cluster occurs beneath a surface fractures field and shows an
approximately N-S trend. Focal mechanism analysis showed that the P axes and T
axes for earthquakes in the western cluster are mostly oriented approximately N-S
and W-E, consistent with stress directions induced by a dike intrusion (MUSUMECI
et al., 2004). The southeastern cluster shows an approximately NNW-SSE alignment
and might be related to an aborted intrusion or active NNW-SSE faults (MUSUMECI
et al., 2004).
For most volcanoes in the world, it is generally difficult to image the detailed
structure of their plumbing systems, due to small numbers of stations and poor
station coverage geometry. DD tomography is less affected by poor station coverage
because it has higher resolution near the source region by using differential arrival
times, as shown in the studies discussed above. Better characterizing the velocity
structure of the volcano and relocating earthquakes would be helpful for better
understanding the volcanic eruption mechanisms and thus predicting volcanic
eruptions and mitigating volcanic hazards.
model uncertainties for both location and most slowness model parameters, but has
lower model resolution for some boundary slowness model parameters. Including the
absolute data with a relatively small weighting into the differential system is useful to
stabilize the system without sacrificing the low model uncertainties and high model
resolution that the differential system can provide.
Methods and applications of three versions of the DD tomography algorithm are
presented: tomoDD, tomoFDD and tomoADD. TomoDD is a local-scale version that
assumes a flat Earth model and uses an ART-PB ray-tracing algorithm. The regional-
scale version tomoFDD considers the curvature of the Earth and uses a finite-
difference travel-time algorithm that can better deal with sharp velocity contrasts and
discontinuities. Both tomoDD and tomoFDD are based on a regularly distributed
inversion grid. In contrast, tomoADD adapts the inversion mesh to match with the
data distribution based on tetrahedral and Voronoi diagrams, such that the ray
sampling densities on inversion mesh nodes are more uniform than the regular grid
case. This approach provides an automated way to take full advantage of the ability
to resolve fine-scale structure in the source region using the differential times. We
then discuss some examples of applying DD tomography to characterize fault zone
structure, image high-resolution structure of subduction zones, and determine the
velocity structure of volcanoes.
There are some issues to be considered in our future work, as follows:
1. Currently we have finished the implementation of the adaptive-mesh scheme on
the flat-model version of DD tomography (ZHANG and THURBER, 2005). We will
implement an adaptive-mesh scheme for the regional-scale version of DD tomo-
graphy.
2. All three versions of the DD tomography algorithm, tomoDD, tomoFDD and
tomoADD, directly solve for P- and S-wave velocities using absolute and differential
arrival times. Knowing 3-D Vp/Vs variations is valuable in order to have a more
complete characterization of the mechanical properties and geological identity of
crust and upper mantle materials (THURBER, 1993). Vp/Vs variations could be
determined directly from the Vp and Vs models if they have essentially identical
quality. It has been observed that in cases where S-wave arrival data are less
numerous and of poorer quality than P-wave data, however, that Vs would be not as
well resolved as Vp, making the interpretation of Vp/Vs variations difficult
(EBERHART-PHILLIPS, 1990; WAGNER et al., 2005). For this reason, it is generally
not appropriate to obtain Vp/Vs ratios by dividing Vp by Vs directly. Alternatively,
Vp/Vs variations can be determined by the inversion of S-P time differences (WALCK,
1988; THURBER, 1993). We plan to add this functionality to the DD tomography
algorithm to directly invert Vp/Vs ratios using both absolute and differential S-P
time differences.
3. The current DD tomography algorithms cannot directly estimate model
resolution and uncertainty due to the substantial amount of data used for the system.
In ZHANG and THURBER (2003), the model resolution and uncertainty were estimated
400 H. Zhang and C. Thurber Pure appl. geophys.,
Acknowledgements
We are extremely grateful to our collaborators for sharing their results with us.
We thank INGV-CT for allowing us to use their Mt. Etna 2001 eruption data. We
especially thank Felix Waldhauser and Steve Roecker for their assistance and helpful
comments in the early stage of tomoDD algorithm development. Megan Flanagan
helped us to realize the tomoFDD algorithm using part of her code. Certain figures
were constructed using GMT (WESSEL and SMITH, 1991). This material is based upon
work supported by National Science Foundation grants EAR-9814192 (Continental
Dynamics), EAR-0125164 (Geophysics), EAR-0346105 (EarthScope), and EAR-
0409291 (Geophysics).
REFERENCES
ASTER, R.C., BORCHERS B., and THURBER C.H., Parameter Estimation and Inverse Problems, (Elsevier
Academic Press, Burlington, MA, 2005) 301 pp.
BARBER, C.B., DOBKIN, D.P., and HUHDANPAA, H.T. (1996), The Quickhull algorithm for convex hulls,
Trans. Mat. Software 22, 469–483.
Vol. 163, 2006 Applications of Double-difference Seismic Tomography 401
BEN-ZION, Y., KATZ, S., and LEARY, P. (1992), Joint inversion of fault zone head waves and direct P arrivals
for crustal structure near major faults, J. Geophys. Res. 97, 1943–1951.
BROWN, J., PREJEAN, S., ZHANG, H., POWER, J., and THURBER, C. (2004), Double difference earthquake
relocation and tomography at Mount Spurr Volcano, Alaska, 1991 to 2004, EOS Trans. AGU 85(47), Fall
Meet. Suppl., Abstract S51A–0140.
DU, W., THURBER, C. H., and EBERHART-PHILLIPS, D. (2004), Earthquake relocation using cross-correlation
time delay estimates verified with the bispectrum method, Bull. Seismol. Soc. Am. 94, 856–866.
EBERHART-PHILLIPS, D. (1990), Three-dimensional P and S velocity structure in the Coalinga region,
California, J. Geophys. Res. 95, 15343–15363.
ENESCU, B., and MORI, J. (2004), Relocations and 3-D velocity structure for aftershocks of the 2000
W. Tottori (Japan) earthquake and 2001 Gujarat (India) earthquake, using waveform cross correlations,
EOS Trans. AGU 85(47), Fall Meet. Suppl., Abstract S54A–05.
FLANAGAN, M.P., MYERS, S.C., SCHULTZ, C.A., PASYANOS, M.E., and BHATTACHARYYA, J. (2000), Three-
dimensional a priori model constraints and uncertainties for improving seismic location, Proceedings of
22nd Seismic Research Review: Technologies for Monitoring the CTBT, 2–15.
HACKER, B.R., ABERS, G.A., and PEACOCK, S.M. (2001), Subduction factory 1, Theoretical mineraology,
densities, seismic wave speeds, and H2O contents, J. Geophys. Res. 108, B12030, doi:2001JB001129.
HOLE, J.A., and ZELT, B.C. (1995), 3-D finite-difference reflection travel times, Geophys. J. Int. 121,
427–434.
JI, S., and WANG, Z. (1999), Elastic properties of forsterite-enstatite composites up to 3.0 Gpa, Geodynamics
28, 147–174.
KATO, A., KURASHIMO, E., HIRATA, N., SAKAI, S., IWASAKI, T., and KANAZAWA, T. (2005), Imaging the
source region of the 2004 mid-Niigata prefecture earthquake and the evolution of a seismogenic thrust-
related fold, Geophys. Res. Lett. 32, L07307, doi:10.1029/2005GL022366.
KLEIN, F.W. (1978), Hypocenter location program HYPOINVERSE, U.S. Geol. Surv. Open File Rep.
78–694, 114 pp.
KOKETSU, K., and SEKINE, S. (1998), Pseudo-bending method for three-dimensional seismic ray tracing in a
spherical earth with discontinuities, Geophys. J. Int. 132, 339–346.
LEE, W.H.K., and STEWART, S.W. (1981), Principles and applications of microearthquake networks, Adv.
Geophys. Suppl. 2, 293 pp.
MONTEILLER, V., GOT, J., VIRIEUX, J., and OKUBO, P.G. (2004), Imaging Kilauea Volcano magmatic system
using an efficient algorithm for double-difference tomography and cross-spectral time delays, EOS Trans
AGU 85(47), Fall Meet. Suppl., Abstract S11D–03.
MENKE, W., (1989), Geophysical Data Analysis: Discrete Inverse Theory, revised edition, vol. 45 of
International Geophysics Series (Academic Press, San Diego) 289 pp.
MENKE, W., and SCHAFF, D. (2004), Absolute earthquake locations with differential data, Bull. Seismol. Soc.
Am. 94, 2254–2264.
MICHELINI, A., and LOMAX, A. (2004), The effect of velocity structure errors on double-difference earthquake
location, Geophys. Res. Lett. 31, L09602, doi:10.1029/2004GL019682.
MINKOFF, S.E. (1996), A computationally feasible approximate resolution matrix for seismic inverse
problems, Geophys. J. Int. 126, 345–359.
MIYAZAWA, M., and KATO, M. (2004), On interpolation functions in travel-time tomography, Geophys.
J. Int. 158, 169–178.
MUSUMECI, C., COCINA, O., GORI, P.D., and PATANÉ, D. (2004), Seismological evidence of stress induced by
dike injection during the 2001 Mt. Etna eruption, Geophys. Res. Lett. 31, L07617, doi:10.1029/
2003GL019367.
NAKAJIMA, J., HASEGAWA, A., ZHANG, H., and THURBER, C.H. (2004), Imaging crust, mantle and slab
structure around the Shikoku district, Japan, by double-difference tomography and its relationship to the
occurrence of nonvolcanic deep tremors, EOS Trans. AGU 85(47), Fall Meet. Suppl., Abstract S51B–0161.
NAKAJIMA, J., MATSUZAWA, T., HASEGAWA, A., and ZHAO, D. (2001), Three-dimensional structure of Vp,
Vs, and Vp/Vs beneath northeastern Japan: Implications for arc magmatism and fluids, J. Geophys. Res.
106, 21,834–21,857.
NOLET, G., MONTELLI, R., and VIRIEUX, J. (1999), Explicit, approximate expressions for the resolution and
a posteriori covariance of massive tomographic systems, Geophys. J. Int. 138, 36–44.
402 H. Zhang and C. Thurber Pure appl. geophys.,
OBARA, K. (2002), Nonvolcanic deep tremor associated with subduction in southwest Japan, Science 296,
1679–1681.
OKADA, T., HASEGAWA, A., SUGANOMATA, J., ZHAO, D., ZHANG, H., and THURBER, C. (2004), Imaging the
fault plane and asperities of the 1995 southern Hyogo (Kobe) earthquake (M7.3) by double-difference
tomography, EOS Trans. AGU 85(47), Fall Meet. Suppl., Abstract S53C–01.
OKADA, T., UMINO, N., MATSUZAWA, T., NAKAJIMA, J., UCHIDA, N., NAKAYAMA, T., HIRAHARA, S.,
SATO, T., HORI, S., KONO, T., YABE, Y., ARIYOSHI, K., GAMAGE, S., SHIMIZU, J., SUGANOMATA, J.,
KITA, S., YUI, S., ARAO, M., HONDO, S., MIZUKAMI, T., TSUSHIMA, H., YAGINUMA, T., HASEGAWA, A.,
ASANO, Y., ZHANG, H., and THURBER, C. (2005), Aftershock distribution and seismic velocity structure in
and around the focal area of the 2004 mid-Niigata prefecture earthquake obtained by applying double-
difference tomography to dense temporary seismic network data, Earth Planets Space 57, 435–440.
PATANÉ, D., CHIARABBA, C., COCINA, O., and GORI, P.D. (2002), Tomographic images and 3D earthquake
locations of the seismic swarm preceding the 2001 Mt. Etna eruption: Evidence for a dyke intrusion,
Geophys. Res. Lett. 29, doi:10.1029/2001GL014391.
PATANÉ, D., GORI, P.D., CHIARABBA, C., and BONACCORSO, A. (2003), Magma ascent and the
repressurization of Mount Etna’s volcanic system, Science 29, 2061–2063.
PODVIN, P., and LECOMTE, I. (1991), Finite difference computation of travel times in very contrasted velocity
models: A massively parallel approach and its associated tools, Geophys. J. Int. 105, 271–284.
RATCHKOVSKI, N.A., and HANSEN, R.A. (2002), New evidence for segmentation of the Alaska subduction
zone, Bull. Seismol. Soc. Am. 92, 1754–1765.
SAMBRIDGE, M., BRAUN, J., and MCQUEEN, H. (1995), Geophysical parameterization and interpolation of
irregular data using natural neighbors, Geophys. J. Int. 122, 837–857.
SHELLY, D.R., BEROZA, G.C., ZHANG, H., THURBER, C.H., and IDE, S. (2004), High-resolution subduction
zone seismicity and velocity structure in Ibaraki, Japan, EOS Trans. AGU 85(47), Fall Meet. Suppl.,
Abstract S43D–04.
TAKEDA, T., KUWAHARA, Y., MIZUNO, T., IMANISHI, K., OKADA, T., ITO, K., WADA, H., and HARYU, Y.
(2004), Crustal structure and micro-seismic activity along the Atotsugawa Fault system, central Japan, EOS
Trans. AGU 85(47), Fall Meet. Suppl., Abstract S13D–1103.
THURBER, C.H., Local earthquake tomography: Velocities and Vp/Vs theory, in Seismic Tomography:
Theory and Practise (eds. H.M. Iyer and K. Hirahara) pp. 563–583 (Chapman and Hall, New York
1993)
THURBER, C.H. (1992), Hypocenter velocity structure coupling in local earthquake tomography, Phys. Earth.
Planet. Int. 75, 55–62.
THURBER, C.H. (1983), Earthquake locations and three-dimensional crustal structure in the Coyote Lake
area, central California, J. Geophys. Res. 88, 8226–8236.
THURBER, C.H., and EBERHART-PHILLIPS, D. (1999), Local earthquake tomography with flexible gridding,
Computers and Geosciences 25, 809–818.
THURBER, C., ROECKER, S., ZHANG, H., Baher, S., and ELLSWORTH, W. (2004), Fine-scale structure of the
San Andreas fault and location of the SAFOD target earthquakes, Geophys. Res. Lett. 31, doi:10.1029/
2003GL019398.
UM, J. and THURBER, C.H. (1987), A fast algorithm for two-point seismic ray tracing, Bull. Seism. Soc. Am.
77, 972–986.
VIDALE, J.E. (1990), Finite-difference calculation of travel-times in three dimensions, Geophysics 55,
521–526.
WAGNER, L.S., BECK, S., and ZANDT, G. (2005), Upper mantle structure in the south central Chilean
subduction zone (30 to 36S), J. Geophys. Res. 110, B01308, doi:10.1029/2004JB003238.
WALCK, M.C. (1988), Three-dimensional Vp/Vs variations for the Coso region, California, J. Geophys. Res.
93, 2047–2051.
WALDHAUSER, F. (2001), hypoDD: A computer program to compute double-difference hypocenter locations,
U.S. Geol. Surv. Open File Rep. 01–113, 25 pp.
WALDHAUSER, F., and ELLSWORTH, W.L. (2000), A double-difference earthquake location algorithm:
Method and application to the Northern Hayward fault, California, Bull. Seismol. Soc. Am. 90,
1353–1368.
Vol. 163, 2006 Applications of Double-difference Seismic Tomography 403
WESSEL, P., and SMITH, W.H.F. (1991), Free software helps map and display data, EOS Trans. AGU 72, 441
and 445–446.
WOLFE, C. J. (2002), On the mathematics of using difference operators to relocate earthquakes, Bull. Seismol.
Soc. Am. 92, 2879–2892.
YAO, Z.S., ROBERTS, R.G., and TRYGGVASON, A. (1999), Calculating resolution and covariance matrices for
seismic tomography with the LSQR method, Geophys. J. Int. 138, 886–894.
ZHANG, J., and MCMECHAN, G.A. (1995), Estimation of resolution and covariance of large matrix
inversions, Geophys. J. Int. 121, 409–426.
ZHANG, H., and THURBER, C. (2005), Adaptive-mesh seismic tomography based on tetrahedral and Voronoi
diagrams: Application to Parkfield, California, J. Geophys. Res. 110, B04303, doi:10.1029/2004JB003186.
ZHANG, H., RATCHKOVSKI, N., THURBER, C., and HANSEN, R. (2004), High-resolution seismic velocity
structure of the Alaska subduction zone revealed by double-difference tomography, EOS Trans. AGU
85(47), Fall Meet. Suppl., Abstract S51B–0168.
ZHANG, H., THURBER, C.H., SHELLY, D., IDE, S., BEROZA, G.C., and HASEGAWA, A. (2005), High-
resolution subducting-slab structure beneath northern Honshu, Japan, revealed by double-difference
tomography, Geology 32, 361–364, doi: 10.1130/G20261.
ZHANG, H., and THURBER, C.H. (2003), Double-difference tomography: The method and its application to
the Hayward fault, California, Bull. Seismol. Soc. Am. 93, 1875–1889.
Zhao, D., Hasegawa, A., and Kanamori, H. (1992), Tomographic imaging of P- and S-wave velocity
structure beneath northeastern Japan, J. Geophys. Res. 97, 19, 909–19,928.